Skip to content

gate: clause 3 holds on b11081 by a direct direct-io A/B; the compose .env names 0.34.4 - #395

Merged
glennneuber merged 1 commit into
mainfrom
gate/0344-dio
Sep 27, 2026
Merged

glennneuber merged 1 commit into
mainfrom
gate/0344-dio

Conversation

@glennneuber

Copy link
Copy Markdown

This updates the 0.34.4 promotion record in docs/maxusai/amd-upgrade-gate.md (#391) in two places.

Clause 3 (--direct-io) is now validated directly on b11081. The record said (b), and (c) only by transitivity, because the payload moved and no direct A/B had run. The maintainer asked for one. It is the same check the 0.34.2 promotion ran on b10969 on 2026-09-21:

  • the vision suite, think off, on qwen3.6 and gemma4:31b, with a cold server per model;
  • the promoted image, in production's environment: f16, flash attention on, two slots, two-pass;
  • 0/54 blocks differ between direct-io on and off, with every scored cell equal (0 of 978 and 0 of 986);
  • both arms verified at the runner flag line: on passed --load-mode dio on both runner starts, and off (OLLAMA_IGPU_DIRECT_IO=0) passed it on neither;
  • all 108 blocks finished with valid JSON.

The script's output is pasted verbatim under the clause table.

The compose .env line. The decision said the .env still named 0.32.1, and warned against a compose redeploy. The .env names the 0.34.4 image since MaxusAI/ollama-deployments 55e2fa2. docker compose config matches production, except for three things:

  • label=disable, which has no effect on this host;
  • an added healthcheck;
  • a compose network in place of the default bridge.

Verification

  • check_source_paths.py --changed-since origin/main is clean.
  • The name scan of the diff finds nothing.

amd-server/rocm-gfx1151

🤖 Generated with Claude Code

… .env names 0.34.4

The 0.34.4 promotion recorded clause 3 (--direct-io) under (b), and under (c)
only by transitivity. The direct A/B has now run on the promoted image in
production's environment, the same check as 2026-09-21 on b10969: vision suite,
think off, qwen3.6 and gemma4:31b, a cold server per model. 0/54 blocks differ
(0 of 978 and 0 of 986 scored cells), and both arms are verified at the runner
flag line (dio-on: --load-mode dio on both runner starts; dio-off,
OLLAMA_IGPU_DIRECT_IO=0: on neither). All 108 blocks finished with valid JSON.

The decision also said the compose .env still named 0.32.1. It names 0.34.4 since
ollama-deployments 55e2fa2, and `docker compose config` matches production but
for label=disable (no effect here), a healthcheck, and a compose network.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@glennneuber
glennneuber merged commit 1abc840 into main Sep 27, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant